A number of recent studies show that weather patterns in major financial centers influence stock

index returns and suggest that investor mood influences asset prices. For example, using data
from international stock exchanges, Hirshleifer
and Shumway (2003) show that stock market returns are higher on days when
the weather is sunny, which is presumably when market participants are in
a good mood.1 The findings from this finance literature are consistent with
evidence in the psychology literature, in that individuals misattribute mood induced by weather
as information when making assessments about objects
that should be otherwise unrelated (e.g., sunny days with positive assessments
about the market).2
In this paper, we focus on the economic mechanism through which weather
induced mood affects asset prices. In particular, we investigate how mood
affects the trading behavior of a group of investors that play a key role in
the price formation process: institutional investors. Despite commonly held
assumptions about investor sophistication, there is evidence that institutional
investors and other sophisticated market participants are also susceptible to
cognitive biases.3 Motivated by these findings, we conjecture that weather
patterns would affect the mood of institutional investors, in tum impacting
their trading decisions and stock prices congruently. We test this conjecture
using financial data from multiple sources that are matched to ZIP code-level
weather station data.
Similar to previous studies, we use deseasonalized cloud cover as our investor
mood proxy. In addition, we use survey data from the Yale International Center
for Finance that provide institutional investor perceptions on stock market
investment. Using the survey data, we are able to construct a variety of measures
that capture perceived investor mispricing. We also employ disaggregated data
that contain the daily trades of institutional investors. Finally, using weather
and institutional holdings data, we construct a novel, stock-level measure of
investor mood to assess its impact on individual stock prices. Our ability to
observe the locations of investors allows us to precisely measure the weather
conditions that they experience at each point in time.
Our findings can be summarized as follows. First, using the survey data,
we show that deseasonalized cloud cover increases the likelihood of perceived
overpricing in both individual stocks and the Dow Jones Industrial Average
(DJIA) among institutional investors. Second, using the trade data, we find
evidence consistent with the survey results, showing that institutional investors
with lower exposure to deseasonalized cloud cover exhibit a greater propensity
to buy. The impact of weather on beliefs and trading behavior is significant both
statistically and economically. In particular, a one-standard-deviation increase
in deseasonalized cloud cover increases the likelihood of perceived overpricing
by approximately 3%, which represents more than 13% of the sample mean.
Further, differences in trade imbalances between investors in the top and bottom 10th percentile
in deseasonalized cloud cover represent more than 9% of its
total sample variation.
Third, we assess how our stock-level proxy for investor mood relates to
individual stock returns. Because U.S. stock ownership is generally dispersed
across investors in different locations, tests using cloud cover in a single
location to proxy for mood may be biased due to poor identification. Our
proxy for investor mood uses institutional investor holdings data to incorporate
information about weather conditions faced by different investors in the same
stock. We find evidence that the stock-level mood proxy has a strong effect on
the daily returns of stocks associated with higher arbitrage costs. Finally, we
document return comovement attributable to the investor mood proxy and find
that these patterns are generally short lived.
These empirical findings contribute to the growing theoretical literature that
identifies the channels through which mood affects economic and financial
market outcomes.4 Studies that focus on the weather effect are motivated
by findings in the psychology literature that shows that affective states can
generate mood-congruent biases in risk-based decisions.5 Accordingly, investor
behavior may be subconsciously influenced by factors affecting mood unrelated
to market information and may in turn impact financial market outcomes (Mehra
and Sah 2002; Bodoh-Creed 2013; Kamstra et al. 2014).6 The salience of
cloud cover as a mood-priming device allows us to both verify the impact
of mood on beliefs and trading behavior of institutional investors and trace
price effects directly attributable to institutional investors using our stock-level
mood proxies.
Recent experimental studies also examine how mood-inducing cues related
to weather affect risk preferences and financial decisions of individuals and
document evidence consistent with mood congruency in assessments. For
example, Bassi, Colacito, and Fulghieri (2013) provide experimental evidence
suggesting that cloud cover has a strong impact on measures of risk tolerance.
Similarly, Kramer and Weber (2012) use experimental evidence to establish
a link between seasonal affective disorder (SAD) and risk attitudes. Whereas
these studies use subjects from the general population, professional investors may not necessarily
be subject to the same dynamics.7 We offer direct
examination of how deseasonalized cloud cover impacts institutional investor
perceptions using survey data and provide evidence that investor optimism,
as measured by lower levels of deseasonalized cloud cover, decreases the
likelihood of perceived investor overpricing.
Other previous studies examine the effect of weather on investor trading
behavior. Loughran and Schultz (2004) and Goetzmann and Zhu (2005) use
disaggregated data to identify the impact of weather conditions across different
geographies. Loughran and Schultz (2004) show cloud cover around the firm's
location has little impact on stock trading volume attributable to mood, except
in cases of extreme weather conditions. Goetzmann and Zhu (2005) match
retail investor trade data to cloud cover in five major U.S. cities from 1991 to
1996. They find the impact of cloud cover on investor propensities to buy and
sell is statistically insignificant. However, they find a strong relation between
bid-ask spreads of NYSE stocks and NYSE index returns with New York City
cloud cover. They interpret these findings as evidence of a weather effect on
market makers.8
To our knowledge, our study is the first to directly test for the weather
effect in institutional investor trading behavior. We provide evidence that
investor optimism increases buy-sell trade imbalances. Further, these results
complement our findings on investor perceptions of mispricing.
We provide evidence that investor optimism increases daily returns of
individual stocks associated with higher arbitrage costs. Binding short-sale
constraints may allow for mispricing to persist (Nagel 2005), as these stocks are
more costly to arbitrage. We acknowledge that our estimates may be understated
due to measurement error in the investor mood proxy. Specifically, we cannot
observe the daily holdings of institutional investors, as we rely on the most
recent quarterly snapshots for construction. As a result, the mood proxy may
not incorporate all investors trading in a stock between the reporting dates.
However, we also find evidence of return comovement attributable to investor
mood and find that these effects do not persist beyond one month.
Taken as a whole, our results provide strong evidence on how mood
influences how institutional investors form their beliefs, affecting their trading
decisions. Together with evidence from Goetzmann and Zhu (2005), our
findings suggest that the weather effect is detectable among sophisticated
market participants. More importantly, we demonstrate how investor mood
can influence various aspects of the price formation process.
1. Theoretical Motivation
A number of studies document a robust relationship between stock index
returns and local weather patterns, referred to here as the "weather effect."
Saunders (1993) shows that sky cloud cover over New York City has a strong,
negative association with New York Stock Exchange (NYSE) index returns.
Because his findings appear to relate to the local weather conditions around
the stock exchange, the results are plausibly attributable to factors related to
investor mood rather than to direct effects on firm fundamentals. Hirshleifer and
Shumway (2003) provide additional evidence using an international sample
of twenty-six stock exchanges. They find a similar relation between cloud
cover in the locations of the stock exchanges and the stock market index
returns, indicating that the relation between weather and stock market returns
is pervasive and not a result of unintended data mining.
Despite strong evidence of the weather effect on stock index returns,
establishing plausibility in mood-based explanations relies in part on
distinguishing which group of investors drives the weather effect. Given the
pervasiveness of these effects, the influence of mood should be detectable at
the investor level. In particular, by employing localized proxies for mood of
investors located in different regions, mood-based explanations of the weather
effect imply that we should observe similar patterns between mood and investor
trading behavior.
The use of sunshine, or sky cloud cover, as a proxy for mood is appealing,
as it is difficult to predict accurately and its influence on mood has been
well documented in the social psychology literature. A broad assertion in that
literature is that individuals can misattribute affective states for information,
which, in tum, affects judgment (Cunningham 1979; Schwarz and Clore 1983).
Several studies use sunshine as a priming device to induce positive affective
states and provide evidence that it corresponds with mood-congruent choices
in cases unrelated to clinical depression.9 The evidence has been documented
in a number of settings, including tipping (Cunningham 1979; Rind 1996),
life satisfaction (Schwarz and Clore 1983), and responsiveness to persuasion
(Clore, Schwarz, and Conway 1994).
Prior research also supports the idea that mood can have an intertemporal
effect on decision making. That is, judgment may not only be influenced
by immediate affective states but also by a composite of affective experienced over time.
Affective states can influence how individuals
encode information into memory; when information is subsequently recalled,
evaluation may be biased congruently according to the composite of prior
moods (Fishbein 1963).10
Additionally, the psychology literature provides evidence that emotion can
strongly influence responses when an individual is asked to extrapolate future
events (Johnson andTversky 1983), as well as when asked to make assessments
on abstract objects in the presence of limited information (Clore, Schwarz, and
Conway 1994).11 Professional investors are likely to use information collected
over time to make financial decisions about a particular investment. Therefore,
the composite of affective states over time should be more informative about
how investor mood affects belief formation and trading behavior than the
investor's affective state at a single point in time.
Turning to the clinical literature, studies that examine the impact of
simulated sunlight on seasonal and nonseasonal depression suggest that
repeated treatments over time, ranging from one to five weeks, are required
to produce antidepressant benefits.12 In their seminal study on seasonal
depression, Rosenthal et al. (1984) show that light treatment simulating
sunshine during clear sky days has stronger antidepressant benefits than does
treatment simulating sunshine in fully overcast days. When subjects were
withdrawn from light therapy, they showed that the subject conditions generally
reversed after one week. Similar evidence is in clinical studies on nonseasonal
depression (Goel et al. 2005).
The influence of cloud cover on mood also may be related to seasonal
variation in the number of daylight hours through its regulating role on
sunlight exposure, which is a well-documented factor responsible for symptoms
associated with SAD, as well as milder forms of seasonal depression (e.g.,
"winter blues"). Kamstra, Kramer, and Levi (2003) examine the influence of
seasonal forms of depression on stock market activity. They find strong evidence
that the number of daylight hours has a positive impact on stock index returns
using a broad set of stock exchanges in different locations spanning the Southern
and Northern hemispheres. Their results also support the role of mood on stock
market activity, though such effects are more likely to be related to disruptions in biorhythm due
to changes in seasons.13 Nevertheless, because light therapy
has been shown to be effective in treating both seasonal and nonseasonal forms
of depression, the influence of cloud cover on mood may not be completely a
seasonal phenomenon.
Our empirical design is motivated by these previous experimental and
clinical studies. Specifically, we use cloud cover to prime investor mood and
assess its impact on investor perceptions and trading behavior. Given that
investors plausibly rely on complex sets of information collected over time
for their decisions, we consider the intertemporal effects of cloud cover over
several weeks, rather than on a single day. Additionally, to distinguish factors
related to seasonal variation in cloud cover, we use deseasonalized cloud cover
as the main priming instrument.
2. Data and Summary Statistics
2.1 Main data sources
Our empirical analysis relies on data from several sources. First, institutional
investor survey data are collected from the Investor Behavior Project at Yale
University. Since 1989, questionnaire survey data have been collected on the
perceptions of investors in the United States about stock market investment.14
Approximately 100 professional investors are surveyed per month from January
2005 to February 2007.15 The response-level data include the date when the
respondent completes the survey as well as his or her location. To ensure data
consistency, we only use responses with values to the survey questions used
in the analysis. Altogether, there are 1,543 responses in the sample before they
are matched to the weather data described below.
Second, the weather data are collected from the Integrated Surface Database
(ISD) and are publicly available from the National Oceanic and Atmospheric
Administration Web site (www.ncdc.noaa.gov). The ISD database contains
hourly weather observations from over 20,000 weather stations worldwide,
of which 11,000 are currently active. For each weather station, we are able
to observe location and weather conditions, including sky cloud cover. We
collect hourly value of the weather variables from all U.S. weather stations
from January 1998 to December 2010.
Third, the institutional daily trading data are provided by ANcerno Ltd.
(formerly the Abel Noser Corporation). The sample period is from January 1999 to December
2010. ANcerno is a widely recognized consulting firm
that monitors equity trading costs of institutional investors, such as CalPERS,
Putman Investments, and Lazard Asset Management.16 The client manager
code along with the institutional client code allows for identification of the
investor. Additional fields used in the analysis include the stock historical
CUSIP number, trade date, trade direction, quantity of shares traded, and trade
execution price.
There are two unique features of the ANcerno data that are central to the
trade-related tests. First, the ANcerno dataset allows us to observe the true
direction (i.e., buy/sell) for all executed trades, eliminating the need to rely on
the Lee and Ready (1991) algorithm. Second, the ANcerno dataset provides the
identities of the institutions, allowing us to hand-collect the ZIP codes of their
locations. Our analysis uses only institutions located in United States, which
represents approximately 80% of all trades in the database.
Fourth, we obtain the 13(f) institutional holdings data from Thompson
Reuters for the 1999:Q1 to 2010:Q4 sample period. The data provides quarterly
snapshots of investor portfolio positions. The ANcerno database provides
trading information for only a subset of the investors in the 13(f) data. We
are able to hand-collect the ZIP codes of the institution's headquarters using
the Nelson's Directory of Investment Managers.
Other data sources used in the analysis are as follows. Stock characteristics
are collected from the Center for Research in Security Prices (CRSP). Only
common stocks (share code of 10 or 11) from January 1999 to January 2010
are included in the analysis. The county-level estimates of median household
income and population are obtained from the Bureau of Economic Analysis
(BEA).
Figure 1 presents the geographical distribution of institutional investors
represented in the survey (panel A), trade (panel B), and holdings (panel
C) datasets at the county level. Shaded counties correspond with regions in
which at least one investor is represented. All three panels convey considerable
geographical heterogeneity in each of the datasets. Panel A shows that the
survey respondents are geographically well represented and are more heavily
represented in regions with greater population. Panel B shows a similar pattern
within the trade dataset, though this geographical distribution is relatively
sparse relative to the survey data. This pattern is not surprising, given the limited
number of investors in the trade database. Panel C shows that the holdings data
have greater heterogeneity than do the trade data, which is again expected
because investors in the trade data represent only a subset of investors in the
holding data.
2.2 Variable construction
2.2.1 Deseasonalized cloud cover measures. The primary variable of
interest from the weather data is the hourly sky cloud cover, which takes on
integer values from zero (sky clear) to eight (full cloud cover). Similar data are
used by Hirshleifer and Shumway (2003) and Goetzmann and Zhu (2005). The
survey and trade data are merged with the weather data using criteria based
on geographical distance. Specifically, using the location coordinates of each
station and investor, we calculate their distances based on the haversine distance
formula.17 The average daily sky cloud cover is calculated using hourly values
from 6 a.m. to 12 p.m., for each date and weather station, the time period
during which investors are most likely to observe outdoor weather conditions.
For each ZIP code, the average daily sky cloud cover is calculated using values
from all weather stations within a 50-kilometer radius of the investor's ZIP
code centroid.18
Because mood may have an intertemporal effect on individuals, we compute
rolling averages of the cloud measure in each ZIP code based on information
from x days before to the response or trade date. Because the average amount of sunlight is a
decreasing, convex function in sky cloud cover, the sky cloud cover
measure used in the analysis (SKC) is defined as the natural logarithm of one
plus the rolling average of the ZIP code-level sky cloud cover.19 Additionally,
we require that each institutional investor has at least one matched weather
station included in the analysis. We note that measurement error in observed
cloud cover is likely to bias our tests from finding any effect.
To adjust for seasonality in the SKC measure, we consider two approaches.
First, we calculate seasonal cloud cover as the average daily cloud cover
for the same month over the entire sample period. Seasonal SKC is the
natural logarithm of one plus the seasonal value and can be used to control
for seasonality when assessing the impact of SKC in the tests. Second,
deseasonalized SKC, or DSKC, is defined as the difference between SKC and
Seasonal SKC.20 Including DSKC measure in place of SKC and Seasonal
SKC in an OLS regression model is equivalent to constraining the model
coefficient on Seasonal SKC to be equal to the negative of the SKC coefficient.
Because this representation may be too strong, we estimate the model using
both approaches.
2.2.2 Investor perceived mispricing measures. We begin the empirical
analysis by focusing on responses from the survey related to perceived
mispricing. To directly assess whether investors believe that stocks are
overpriced (or underpriced) relative to fundamentals, we examine the
association between the weather measures and responses to the question, "Stock
prices in the United States, when compared with measures of true fundamental
value or sensible investment value, are: (a) Too low, (b) too high, (c) just
right, and (d) do not know." An indicator variable, TooHigh is constructed,
which takes the value of one to a response of "too high" and zero otherwise.
Similarly, TooLow takes the value of one to a response of "too low" and zero
otherwise.
Next, we construct a continuous measure of perceived mispricing using
reported values of what the respondent believes to be the intrinsic level of
the DJIA.21 The responses provide estimates of the investors' perceptions on
the intrinsic value of the DJIA and can be compared with the actual DJIA
level around the survey date. A measure can be constructed representing the
percentage mispricing, or %DJIAMisPrc, defined as the natural logarithm of
the ratio between the survey response and the average DJIA level over the past seven days.
Lower values of %DJIAMisPrc are associated with relatively
greater overpricing.22
In the sample of investors examined, we conjecture that pessimism should be
positively related to perceived overpricing, as most of the survey respondents
are more likely to be limited in their ability to short stocks. However, there
are several key limitations affecting inferences with the TooHigh and TooLow
variables. Tests on those measures must assume that the respondents apply
similar criteria to assess mispricing. However, this assumption is likely to be
too strong. To address this issue, we construct similar binary variables from the
%DJIAMisPrc measure. Specifically, we code DJIATooHigh as a value of one
if %DJIAMisPrc is below the bottom sample quartile. Similarly, DJIATooLow
is coded with a value one if %DJIAMisPrc is in the top sample quartile.
The DJIATooHigh and DJIATooLow measures are based on a stock index
rather than on individual stocks, making it more likely that the respondents
are basing their responses on their perceptions of market trends rather than in
the context of their own portfolio holdings. We assess these five measures of
perceived mispricing to validate whether cloud cover is positively associated
with bad mood states.
2.2.3 Investor trade measures. We construct two measures based on the
institutional trade data. For both measures, we aggregate trades across all
investors for a particular stock within the same ZIP code. First, we construct
a measure of investor-level buy-sell imbalance, or Investor BSI Ratio, that
aggregates all buy and sell trades across stocks and is defined as the difference
between the daily dollar buy and sell volume across all investors in the dataset
in a particular ZIP code, scaled by the total dollar volume in the same date and
ZIP code. When there is no trade on a particular date, Investor BSI Ratio takes
a value of zero. Second, we construct an analogous measure of daily buy-sell
imbalance by investor and stock, or Stockinvestor BSI Ratio. The measure is
defined as the daily, buy minus sell dollar volume scaled by the total dollar
volume for the same date, stock, and ZIP code. Because we cannot observe
daily positions, we only use dates where investors trade in the stock.
We aggregate trades at the ZIP code level given that the trade and weather
data are matched using investor ZIP code. Idiosyncratic trading behavior related
to other factors may serve to increase noise in the BSI measures unrelated to
geographical factors. To alleviate these issues, the analysis restricts the sample
to only ZIP codes with at least three investors at each point in time and with
at least half of the sample period available for the trade data. In addition, we
remove extreme observations associated with ZIP codes in the top and bottom
1% of the total dollar buy-sell volume over the entire sample.
2.2.4 Stock-level mood proxies. We construct a stock-level measure of
investor mood, or StockDSKC, measured approximately as the average DSKC
across locations of institutional investors based on their portfolio holdings.
Institutional investors in stock i are identified using the most recent 13(f)
holdings data. The weather data are linked to the ZIP code of the institution's
location. StockDSKC is calculated as the logarithm of one plus the difference
between the average SKC and Seasonal SKC of all institutional investors in
stock /on date /. StockDSKC is updated daily, using the most recent holdings
data.23
2.2.5 Seasonal affective disorder proxy. In addition to the seasonally
adjusted cloud cover measure, we also condition the tests on seasonal factors
that may also affect mood. Our proxy for seasonal affective disorder, or SAD,
corresponds with the number of nighttime hours, which can be approximated
using a mathematical representation that takes into account the time of year and
the locational coordinate. SAD equals zero during spring and summer months
and the number of nighttime hours minus twelve during the other months. We
use the coordinate centroids of each investor's ZIP code location to calculate
the SAD measure, using the procedure described by Kamstra, Kramer, and
Levi (2003).24 We note that the measure contains significantly less variation
in our sample than that of Kamstra, Kramer, and Levi (2003), given that our
sample focuses on the contiguous United States. On the other hand, they use
an international data sample. Additionally, the length of the sample period in
our data sources is also significantly smaller.
2.3 Summary statistics
Table 1 provides the summary statistics for the weather, survey, and trade
variables used in the analysis. Panel A describes the ZIP code level
deseasonalized sky cloud cover measures across different estimation windows.
The estimation window of x days calculates average DSKC using data from
days t — x to t. Because there is little theory to help guide the selection of the
estimation window used in the analysis, the comparison helps in pinning down
an appropriate length that yields well-behaved estimates.
As the estimation window increases, the sample average decreases in absolute
magnitude. The sample standard deviation also decreases monotonically,
following a convex pattern in the length of the estimation window. In particular,
the sample standard deviation of the measure that uses a one-day window,
which is most similar to the weather measure used by Hirshleifer and Shumway
(2003), is 0.695. The sample standard deviation decreases considerably up to the two-week
window (0.263) and is relatively stable when using a four-week
window (0.223).
Panel B describes the primary variables used in the survey-based tests.
The binary response variables are displayed in the first four rows. In our
sample, slightly more investors indicate that stock prices are generally too high
(19.5%) than too low (14.8%) relative to fundamentals. By construction, the
DJIATooHigh and DJIATooLow variables have a sample mean of approximately
25%. The volatility of daily DJIA returns over the past 30 days is represented
in points.
Panel C provides descriptions on the trade sample. The DSKC measure is
constructed over the trade sample and has a lower sample mean and standard
deviation than those calculated over the survey sample. The Investor and
Stockinvestor BSI ratios are slightly positive in the sample period.
Finally, panel D provides statistical descriptions on the holdings sample. The
changes in stock-level DSKC measures have a lower sample mean than either
the survey or trade data samples. The sample standard deviations slightly half
of the size less than that of the DSKC in the trade sample, which in part is due
to averaging DSKC in stocks with a larger number of investors.
Weather-Induced Mood and Perceived Mispricing
The social psychology literature provides evidence that individuals provide
mood-congruent assessments on objects unrelated to the cause of their affective
states (Cunningham 1979; Schwarz and Clore 1983). In particular, Schwarz and
Clore (1983) find that sunny days are associated with relatively more optimistic
assessments. Recent experimental evidence also shows that weather conditions
can influence risk attitudes through its impact on mood (Bassi, Colacito, and
Fulghieri 2013).
In this section, we analyze the survey data where institutional investors
are asked a series of questions related to their opinions on stock market
investment. Our tests focus on perceived investor mispricing. Investors may
view stocks as mispriced for a variety of reasons. While risk preferences across
professional investors may be relatively homogenous, variation in weather
patterns may generate mood-induced biases in investor beliefs about underlying
fundamentals. In turn, its impact on beliefs may also affect their trading
decisions, which we examine on trade data in subsequent tests. Because the
survey does not draw attention to weather conditions during the interview,
we view the analysis as a field experiment where deseasonalized weather
conditions are randomly assigned to subjects.
3.1 Univariate estimates
Table 2 presents the Pearson correlation coefficients amongst the variables
used in the regression models.25 Panel A presents the results for the dependent
variables and the conditioning variables. Panel B presents the results for
the dependent variables and deseasonalized SKC measured over different
estimation windows, defined similarly to panel A of Table 1. The correlation
coefficients are accompanied with p—values, and are displayed in parentheses. In panel A of
Table 2, the correlation coefficients amongst the survey
variables related to mispricing suggest internal consistency in the responses.
The correlation between TooHigh and DJIATooHigh is 67.8% (p-value <0.001),
and both are strongly, negatively associated with the TooLow, DJIATooLow, and
%DJIAMisPric measures, respectively.
We next examine how other variables correlate with the mispricing measures.
The volatility of DJIA returns has a positive and statistically correlation with measures related to
perceived underpricing, and the correlation
coefficient is higher with the DJIATooLow measure. The long-term corporate
earnings growth rate estimates are signed consistently with received wisdom,
though only the DJIATooLow correlation is statistically significant. Portfolio
size does not have a statistically significant correlation coefficient with
any of the mispricing measures. Neither levels nor changes in the county
level characteristics appear to be consistently correlated with the mispricing
measures.
In panel B, the correlation coefficients between the mispricing measures
and the DSKC are not statistically significant when using estimation windows
of less than one week. When expanding the estimation windows to two
weeks and beyond, the correlation coefficients become statistically significant
for TooHigh, %DJIAMisPrc and DJIATooHigh. Both the TooHigh and
DJIATooHigh coefficients are positive, while the %DJIAMisPrc coefficient
is negative, as expected. Additionally, the DJIATooHigh coefficient is
considerably larger than that of TooHigh. Finally, the correlation coefficients
that use TooLow and DJIATooLow remain statistically insignificant, however,
and only DJIATooLow has the predicted sign. Again, we cannot observe the
criteria investors use to assess what is "too" high or low in the TooHigh and
TooLow measures, respectively. In contrast, the DJIATooLow measure uses
consistent criteria that we define, and so may explain differences in the signs
between the two coefficients.
These results raise some interesting questions. First, the relation between
the coefficients and the length of the DSKC estimation window may be due to
measurement error. The respondent may take more than one day to fill out the
survey, so that the relevant estimation window for the DSKC measure may be
larger than just the day the survey was completed. Using this interpretation,
the results would suggest that the respondent may have taken more than two
weeks to ponder the survey questions, which does not appear to be plausible.
On the other hand, investors may rely on information that has already been
collected over time to determine their responses. The DSKC measure using
longer estimation windows may better correspond with the composite affective
reaction for the periods over which the relevant information is collected.
Additionally, the results are also consistent with findings from clinical studies on
light therapy and nonseasonal depression, showing that antidepressant benefits
of simulated sunlight take hold only after repeated exposure over long treatment
periods.
Second, while the correlation coefficients are large and statistically
significant for the overpricing measures, the coefficients are statistically
insignificant on the underpricing measures. An alternative interpretation of the
findings on the TooHigh measure is that deseasonalized sunshine is positively
related to responses of "too low" or "just right". Bad moods may induce
individuals to examine information with greater scrutiny (Schwartz 1990;
Petty, Gleicher, and Baker 1991), so that good moods may not necessarily result in bullish
responses over non-critical ones. The lack of statistically
significant results associated with the TooLow measure is consistent with this
interpretation. A similar pattern holds for the DJIATooHigh and DJIATooLow
results.
These findings suggest that longer estimation windows used to construct
the cloud cover measures may be important to capture mood effect on the
mispricing measures. For the remainder of the analysis, we include only DSKC
measured over a two-week estimation window given our findings from Table 2,
and are consistent with the clinical studies discussed in Section 1. We next
determine the robustness of the univariate results by conditioning on other
factors that may also impact the survey responses.
3.2 Perceived mispricing regression specification
The OLS regression models of perceived mispricing are specified as follows:

Yijt = b0 +b*DSKQit + b * XM+eiit

The dependent variables in the regression models are TooHigh, TooLow,

%DJIAMisPrc, DJIATooHigh, and DJIATooLow. DSKC, t is constructed as
a 14-day rolling average using information up to date /of zip code-level cloud
cover. Xj,t represents a vector of other explanatory variables for respondent i
at date /.
We include the following conditioning variables in the models. First, proxies
for local economic conditions are included, such as county-level population,
changes in population, county-level median income, and changes in median
income. This choice is motivated by the evidence in Coval and Moskowitz
(1999) document local bias in investment choices of institutional investors.
Economic conditions may affect the future prospects of local investment
opportunities, which in turn may influence investor opinions on stock market
investment.
Second, professional investors with greater assets under management may
have systematically different opinions about equity market investment, and
so the natural logarithm of one plus the size of the respondent's investment
portfolio is included in the models. Third, large, recent fluctuations in stock
prices may also influence responses, as periods of greater equity market
volatility may negatively bias the responses. Accordingly, the natural logarithm
of the sample volatility of the DJIA index returns over the past 30 days is
included as a conditioning variable. Fourth, the respondent's estimate of the
long-term corporate earnings growth is also included, as it may systematically
influence the mispricing measures. Finally, we also include the SAD measure
used in Kamstra, Kramer, and Levi (2003) to further distinguish from seasonal
effects on mood.
For the binary dependent variables, the test coefficients may be influenced
by the functional form of the estimator. To assess its effect, we also estimate the models using a
probit estimator, which takes the form:
P(Yi,t|Ziit)=0(Zu)

Here, Y represents the binary variables associated with perceived mispricing,

and Zj,t = co+ci *DSKCi>t+c*Xi,t. As with the OLS specifications, the
residuals across time and region are unlikely to be uncorrelated, biasing
downward the standard error estimates. To address these issues, only two-way
clustered standard errors on the zip code- and date-levels are reported.26
3.3 Perceived mispricing regression estimates
Table 3 presents the estimates. Panel A presents the OLS model estimates from
all the mispricing models. Panel B presents the probit model estimates using
only the binary dependent variables. Only marginal effects are reported in panel
B to facilitate comparison with the OLS estimates.
Panel A shows that the OLS coefficients on DSKC are consistent with the
univariate results, even after inclusion of the conditioning variables. The DSKC
coefficient in model 1 is positive and statistically significant (estimate = 0.073,
?-val ue = 1.87). In other words, higher deseasonalized cloud cover increases
the likelihood that investors believe that stock prices are "too high" relative to
fundamentals. When DSKC is replaced by SKC and Seasonal SKC in model
2, similar results obtain: the SKC coefficient is 0.087 (t-value = 2.25), while
the Seasonal SKC coefficient is statistically insignificant (estimate = —0.039,
t-value = —0.73).
In contrast, the DSKC coefficient in the TooLow model is not statistically
significant (estimate = 0.029, t-value = 0.89). The DSKC coefficient in the
%DJIAMisPrc model is negative (estimate = —0.025, t-value = —2.14), and
shows that deseasonalized cloud cover decreases the continuous measure of
perceived underpricing. However, most of the explanatory power of DSKC
appears to be driven by responses associated with overpricing. The DSKC
coefficient in the DJIATooHigh model is 0.120 (t-value = 3.04), while that of
the DJIATooLow model is —0.045 (f-value = —0.84). The coefficients on the
conditioning variables are statistically insignificant for the most part, and their
signs are consistent with the univariate results from panel A of Table 2.
We next repeat the analysis using probit estimators. The results are reported
in panel B, and are quite similar. The marginal effect of DSKC in the TooHigh
model is slightly larger (estimate = 0.079, z-value = 1.94), and the result is
similar in the DJIATooHigh model (estimate = 0.125, z-value = 3.07). To
assess the economic magnitudes, a one-standard-deviation increase in DSKC
translates to an increase of 0.021 and 0.033 in TooHigh and DJIATooHigh and
represents 10.6% and 13.1% of the sample means, respectively.
3.4 Perceived mispricing regression estimates: Robustness checks
One concern with the estimates in Table 3 is that they are susceptible to
omitted variable biases that may be driven by seasonal or interregional factors.
While the DSKC measure is constructed to purge seasonal variation, the
seasonal adjustments may be insufficient so that time effects may overstate
the effect of DSKC. Interregional factors invariant of time may also overstate
the effect of DSKC, so that the results may be driven by a limited number of
regions.
To assess the impact of these potential biases, we re-estimate the models from
Table 3 after including state (panel A) or year-quarter (panel B) fixed effects,
which should decrease the statistical power of the tests. Because of sample
size limitations, using regional and time variables with greater granularity is
infeasible. Finally, because of the inclusion of these fixed effects, we only
estimate the models using OLS estimators, given the well-known, incidental
parameters issues associated with fixed effect probit estimators.
In panel A, we show that the results from the state fixed effects models are
similar to those in Table 3. All models include the conditioning variables from
the models in Table 4, though the coefficients are not reported to conserve space.
In the TooHigh model, the DSKC coefficient remains positive but becomes
statistically insignificant at the 10% level (estimate = 0.067, f-value = 1.64).
However, the SKC coefficient remains positive and statistically significant
(estimate = 0.105; f-value = 2.54). The DSKC coefficients remain statistically
significant in the %DJIAMisPrc (estimate = —0.021; f-value = —1.85) and
DJIATooHigh (estimate = 0.112; f-value = 2.61) models. The DSKC coefficients
are statistically insignificant for the TooLow and DJIATooLow models. The
results for the SKC coefficients are similar. Altogether, the estimates are slightly
smaller in absolute magnitude than those in Table 3, though most remain
statistically significant.
Similar results obtain for the year-quarter fixed effects models in panel
B. In the TooHigh model, the DSKC coefficient is positive and statistically
significant (estimate = 0.066; f-value = 1.69), as is the SKC coefficient (estimate
= 0.068; r-value = 1.77). The DSKC coefficient in the %DJIAMisPrc model
remains negative but statistically insignificant at the 10% level (value =—0.017;
f-value = — 1.54), though the SKC coefficient is statistically significant (estimate
= —0.020; r-value = —1.73). The DSKC coefficient in the DJIATooHigh model
is positive and statistically significant (estimate = 0.093; f-value = 2.65). As
with the state fixed effects models, the DSKC coefficients are statistically
insignificant for the TooLow and DJIATooLow models.
Finally, we consider interaction terms in the perceived mispricing models
related to DSKC and SAD. Because cloud cover is also negatively related to
sunlight exposure, which, in the clinical literature, has been documented to
be associated with seasonal affective disorder, it is plausible that cloud cover
may amplify the effects of daylight on individual mood. We find that SAD
coefficients across all the models we estimate are statistically insignificant at the 10% level, and
we posit that the lack of variation in nighttime hours in our
data sample may be the explanation.
To evaluate these views, we include DSKC and SAD interaction terms in the
OLS models from Table 3, and the results are displayed in Table A. 1 in the
Online Appendix. The DSKC coefficient remains statistically significant for
the TooHigh, %DJIA MisPrc, and DJIA TooHigh models. In these models, the
interaction term coefficients are of consistent sign as our conjecture, though
only the one in the TooHigh model is statistically significant at the 10% level.27
We conclude, as before, that the tests may not have sufficient statistical power
given the lack of variation in the number of nighttime hours in our sample and
leave further tests to future research.
4. Weather-Induced Mood and Institutional Investor Trading
The results from the previous section show that DSKC has a strong impact on
perceived mispricing among institutional investors and imply that DSKC may
also impact trading behavior. We evaluate this channel by constructing tests
using daily institutional investor trades from the ANcerno database. Positive
(negative) trade imbalances relate to institutional investors that are net buyers
(sellers) for an observation date. The daily trade data are aggregated to the
ZIP code level and are used to construct the Investor and Stockinvestor BSI
measures. DSKC is constructed identically to the procedure described in the
survey-based tests.
4.1 Mood and trading behavior: Univariate results
We begin by using paired t-tests to assess whether Investor BSI is systematically
higher on sunny relative to cloudy regions across time. We define sunny
(cloudy) regions based on whether DSKC for a particular investor is in the
top (bottom) xth percentile of the sample for each date. The paired t-tests
assess the statistical significance of the difference in group means of Investor
BSI across sunny and cloudy regions conditional upon x. We estimate cases in
which x takes on values of 50,33,25, and 10, respectively. For example, when
x = 10, the paired t-test assesses the difference in group means of Investor
BSI across investors in the top and bottom 10th percentile. In this manner, we
are able to verify whether the difference in the group means increases using
relatively more extreme DSKC values.
Table 5 displays the results. The thresholds used to define the investor groups
are displayed in the first column. The average Investor BSI for the cloudy and
sunny groups is displayed in Columns 2 and 3, respectively. The difference
in the group means is displayed in the fourth column. The differences are
accompanied by their standard errors and are displayed in parentheses. When defining the groups
by the top (bottom) 50th percentile on DSKC for
each sample date, the difference in Investor BSI is negative and statistically
significant (estimate = —0.013; f-value = —3.71), as expected. Tightening
the threshold serves to increase the group differences and appears to be
primarily driven by increasing values of average Investor BSI in the sunny
groups. For cloudy (sunny) groups based on the top (bottom) 10th percentile
on DSKC, the difference in the group means is —0.046 (/-value = —5.48),
which is approximately 9.2% of the sample standard deviation for Investor
BSI. Altogether, these univariate tests show that DSKC has a negative impact
on investor propensities to buy, as predicted. Additionally, the effect becomes
pronounced when tightening the thresholds for the DSKC groupings.
4.2 Multivariate BSI regression specification
Next, we assess the robustness of the univariate results by using panel
regression models to condition variation related to cross-sectional and time
series determinants of Investor BSI. The model estimated over the full sample
is specified as follows:
Investor BSIj^do+d! *DSKCz>[t_i4,t]+d*WZit+ez,t.
For ZIP code z and date t, Wztrepresents a vector of variables that allows
us to isolate the source of the explanatory power of DSKC in the models.
The times-series variables include dummy variables for whether the sample
date is in January or on a Monday, as done by Goetzmann and Zhu (2005),
and the SAD measure, which is constructed using the ZIP code of the trading
institution's location. As with the survey-based tests, we also control for county
level economic variables based on the institution's location. Finally, we also
exploit within-ZIPcode and within-date variation in the DSKC variable through
inclusion of fixed effects as robustness checks after presenting the results on
the baseline models. Table 6 displays the estimates from these models.
In the results reported in Table 7, the dataset is disaggregated further into a
three-level panel, where the data is represented on the ZIP code (z), date (f),
and stock (j) level. The dependent variable of those models is Stockinvestor
BSI. Additionally, unlike the Investor BSI measure, because the dataset not reveal the investor's
daily positions, the tests are conditional on whether
the investor trades (e.g., either buys or sells) in the stock for a particular date.
The advantage of the three-level panel is that we can now account for stock
level factors. Specifically, we include in the regression models individual stock
characteristics related to size, which is measured as the natural logarithm
of the stock's market capitalization, and liquidity, which is measured as the
inverse of the stock's share price. We also examine another specification that
includes fixed effects for each stock-date pair, which is expected to significantly decrease the
statistical power of the tests. The fixed effects models limit the
variation in DSKC to within-stock-date groupings, purging the estimators of
any unobservable factors associated with a particular stock at each point in time.
The fixed effects model described above is specified as follows:
Stockinvestor BSIzj t = fo+fi *DSKCz t+f2 *CountyCharacteristcs/ t
+f3*SADz,t + 7^y7j,t*D(Stock=jnDate = t)+ez,j,t.

jt

The fixed effects model does not include a number of variables that are
subsumed by the fixed effects terms, such as the stock characteristics and the
time-series variables. Because the residuals in the three-level panel model are
unlikely to be independent across the panel dimensions, the reported standard
errors in the tables are adjusted for three-way clustering at the ZIP code, date,
and stock levels.
4.3 Investor trading model estimates
The analysis proceeds by estimating the Investor BSI models. As shown in
Table 6, models 1 and 2 represent the estimates without the fixed effects. Models
3 and 4 include ZIP code fixed effects. Models 5 and 6 include date fixed
effects. The odd-numbered models provide estimates on DSKC, whereas the
even-numbered models show estimates on SKC controlling for Seasonal SKC.
In model 1, the DSKC coefficient is negative and statistically significant
(estimate = —0.024; f-value = —2.25), and the SKC coefficient in model 2
is also negative and statistically significant (estimate = —0.023; /-value =
—2.25). Among the conditioning variables, only the Monday dummy variable is
negative and statistically significant. When including the ZIP code fixed effects
in models 3 and 4, the results are similar. The DSKC coefficient remains stable
(estimate = —0.021; /-value = —2.03) relative to the estimates in model 1.
Finally, similar results obtain when we include date fixed effects, though the
DSKC coefficient is slightly higher (estimate = —0.031 ; /-value = —2.28). The
Monday and January dummy variables are dropped from the model, as they
are subsumed by the date fixed effects.
Because the DSKC measure is constructed in an overlapping manner given
the panel data structure, one possible concern is that the standard error estimates
are inconsistent, due to autocorrelated residual terms. We re-estimate the
standard errors for the model coefficients from Table 6 using the procedure
described by Driscoll and Kraay (1998), which accounts for serial and cross
sectional dependency in the regression residuals. The results are reported in
Appendix Table A.2 and are similar to those in Table 6.
Table 7 presents the estimates for the Stockinvestor BSI models. Models 3,
4, 7, and 8 include the stock-date fixed effects, whereas models 5 through 8
also include the other conditioning variables. The DSKC coefficient estimate
in model 1 is negative and statistically significant (value = -0.008; /-value =
— 1.79). When we include fixed effects in model 3, the adjusted R2 increases dramatically, from
0.001% to 52.530%. However, the DSKC coefficient
remains stable and statistically significant (estimate = —0.007; r-value =
—2.34), suggesting that the explanatory power of DSKC on Stockinvestor BSI
is not affected by stock-date factors. When including the control variables in
model 7, the DSKC coefficient estimate remains similar (estimate = —0.007;
t-value = —2.60). Finally, the results are similar when we replace DSKC with
SKC and Seasonal SKC. Overall, our trading regression results are remarkably
robust even when accounting for stock related factors.
In summary, consistent with our predictions, we find that mood as measured
by cloud cover affects institutional trading behavior. Institutions are less
likely to engage in buying activity during relatively cloudier periods. DSKC
negatively impacts Investor BSI across a wide variety of specifications.
Additionally, these relationships are robust to alternative tests that take into
account stock-level factors.
5. Investor Mood, Stock Prices, and Return Comovement
Our results from the tests in the previous sections demonstrate that weather
based proxies of mood impact institutional investor beliefs and trading behavior.
A related question is whether mood also impacts stock prices through its effect
on investors, which is the focus of this section. Our key finding is that stock-level
mood proxies that incorporate deseasonalized cloud cover across institutional
investors, or StockDSKC, negatively impact returns of stocks with higher
arbitrage costs. We also provide evidence of return comovement attributable to
the mood proxies.
5.1 Investor mood and stock prices: Regression specification
Because the trading database only represents a fraction of the entire universe
of institutional investors, our tests in this section utilize 13(f) holdings data
to construct stock-level measures of investor mood. We restrict the sample to
stocks represented in the 13(f) holdings data that can be matched to the weather
data. StockDSKC is expected to have a negative relationship with daily stock
returns.
Hirshleifer and Shumway (2003) posit that weather fluctuations are more
likely to impact the sign of stock returns rather than the magnitude and so we
employ similar tests. In particular, we estimate an OLS regression model in
which the dependent variable of interest is a binary variable, or PosRet, which
takes value one for the positive stock return days, and zero otherwise.

PosReti>t = g0+gi * StockDSKCj t+g * Xi> t+£i,t

Here, Xj,t represents a set of explanatory variables that includes the natural
logarithm of the market capitalization from the end of the previous quarter, the
proportion of shares held by institutional investors in the stock, the inverse of
the share price from the end of the previous quarter, a Monday dummy variable, a January
dummy variable, and a composite SAD measure that is constructed
similarly to StockDSKC. The coefficient on the StockDSKC variable is expected
to be negative, or g] <0.
Several econometric issues are expected to bias our tests away from our
predictions. First, the explanatory power of StockDSKC is expected to vary
according to costs related to arbitrage. Mispricing arising from trades motivated
by investor mood is expected to be quickly corrected by other investors in the
absence of short-sale constraints. However, when these constraints are binding,
the effects of mispricing are expected to persist (Nagel 2005). We measure
arbitrage costs, or ArbCosts, associated with binding short-sale constraints as
the inverse of one plus the proportion of shares held by institutional investors
in a particular stock, so that higher ArbCosts corresponds with stocks that are
more likely to have binding short-sale constraints. We expect the weather effect
to be pronounced in stocks with higher ArbCosts and estimate effects across
subsamples based on ArbCost rankings accordingly.
Second, the frequency of the holdings data does not allow us to observe
investor positions within each quarter. In particular, we cannot identify investors
that may actively trade in a stock but hold no position at the SEC filing dates.
Measurement errors related to the observation frequency are unlikely to be
systematic, which could lead to attenuation bias on the coefficients in our tests.
Additionally, investors with short positions are not accounted for in StockDSKC,
so that pessimistic investors may not be represented in the measure.
5.2 Investor mood and stock prices: Stock return regression estimates
Table 8 displays the OLS results from the PosRet models. Panel A reports the
estimates when we include only the weather variables in the models, whereas
panel B reports the estimates when all the conditioning variables are included.
The coefficient estimates for the conditioning variables are not reported to
conserve space.
The first two models report the pooled model estimates. The remaining eight
columns report the model estimates on sample splits based on the ArbCosts
rankings. Models 3 and 4 are estimated for the sample in which ArbCosts is
the lowest 50th sample percentile. Models 5 and 6 are estimated on the sample
in which ArbCosts is between the 50th and 80th sample percentiles. Models 7
and 8 are estimated on the sample in which ArbCosts is between the 80th and
95th sample percentiles. Models 9 and 10 are estimated on the sample in which
ArbCosts is between the 95th and 100th sample percentiles.
The first row displays the estimates on the StockDSKC variable. In panel
A, the StockDSKC coefficient in the pooled sample is statistically insignificant
(estimate = —0.003; /-value = —0.14) with only the weather variables. However,
the results from the regression models using subsamples based on ArbCosts
rankings suggest that the weather effect is concentrated among stocks with
higher arbitrage costs. For stocks with ArbCosts in the 95th to 100th sample
percentile range, StockDSKC is negative and statistically significant = —0.029; f-value = —
3.03). The coefficients before that point are relatively
higher but mostly statistically insignificant. Using StockSKC and Seasonal
StockSKC instead yields similar results. When we include the conditioning
variables, panel B shows that the StockDSKC coefficients attenuate somewhat
but remain statistically significant.
The StockDSKC coefficient is statistically significant in stocks within the top
sample quintile based on ArbCosts. The average proportion of shares held by
institutional investors within those stocks ranges from 2% to 20%. Although
we cannot construct mood proxies for stocks not represented in the 13(f)
holdings data, these stocks are expected to be characterized with relatively
higher arbitrage costs. Accordingly, it may be reasonable to expect a detectable
mood effect in stocks with no institutional ownership.
5.3 Investor mood and stock prices: Robustness checks
We consider several alternative specifications to assess the robustness of the
OLS results. First, using probit instead of OLS regression models, we estimate
similar tests in Table 9, which is formatted similarly to Table 8. As before,
only marginal effects are reported for comparability with the OLS estimates.
The marginal effects are virtually identical to the OLS estimates, though the
standard error estimates are slightly larger.
Second, we re-estimate the standard errors for the OLS estimates to correct
for serial dependency in the residual terms using the procedure described by
Driscoll and Kraay (1998). Table A.3 from the Online Appendix reports the
results and shows that the Driscoll-Kraay standard errors are generally smaller
than those obtained using two-level clustering.
Third, we examine whether our results are sensitive to the measure of
arbitrage costs used. In untabulated estimates, we obtain similar results when
we redefine ArbCosts using other proxies, such as market capitalization or
idiosyncratic volatility, though the results are slightly weaker.
Finally, we consider whether our results can be explained in part by the
effect of mood fluctuations of market makers in NYSE stocks. Goetzmann
and Zhu (2005) show that New York City cloud cover negatively impacts
daily NYSE index returns and most of explanatory power is concentrated
on days that experience the largest change in spreads in S&P 100 stocks. In
untabulated results, we find similar results when we exclude NYSE stocks. We
also explore alternative specifications that include New York City cloud cover
as an additional conditioning variable and find that it has little impact on our
findings.

5.4 Investor mood and return comovement

The results from the previous section confirm that the investor mood proxy
has a negative impact on daily stock returns. Using StockDSKC as our proxy
for investor mood, we next examine whether investor mood can induce return comovement. We
also examine the duration of these effects across different
time horizons.
We adapt the return comovement tests from Green and Hwang (2009) and
Kumar, Page, and Spalt (2013) for the analysis. Because we are interested in
relatively shorter investment horizons than what is considered in those studies,
we use daily stock returns data for estimation. First, we construct two portfolios
based on daily StockDSKC rankings, using only stocks either in the lowest or
in the highest 30th sample percentile for each date. The resulting two portfolios
are defined as low StockDSKC (Optimistic) and high StockDSKC (Pessimistic).
For each portfolio j, Moodj is the portfolio return in excess of the risk-free
rate and is value weighted and rebalanced daily.28
Second, the coefficient on each Mood portfolio is estimated for each stock
using rolling, time-series regressions that include the Mood portfolios, the Fama
and French (1993) factors (MKTRF, HML, and SMB), and the Carhart (1997)
momentum factor (UMD). If stock i is included in either of the Mood portfolios
at any point in time, the Mood portfolio is reconstructed to exclude stock i to
avoid any mechanical relation. The model takes the following form
Ri.t-Rf,t = A)+ ^ ßj Moodjt+/ImktrfMKTRFt

j e [Pessimistic,
Optimistic}
+ßuMLliMU + ßsMB SMBt + /IumdUMDi+£j t

The comovement measures are the coefficients from the Mood portfolios.
We use forward-looking 30- and 90-day estimation windows for the rolling
regression models, each resulting in a panel of parameter estimates based on
overlapping data. For inclusion in the sample, we require that each estimate is
calculated using a sufficient number of returns observations.
Finally, we compare the comovement estimates across stocks based on the
StockDSKC rankings. Specifically, we calculate the value-weighted average
of each portfolio loading for quintile groups based on StockDSKC rankings
for each date. We predict that j0pessimisticwill be higher for stocks with higher
StockDSKC values. Likewise, we predict that ^optimisticwill be lower for stocks
with higher StockDSKC values. We calculate these estimates for both the 30
day and 90-day estimation windows. Because of the overlapping estimation
windows used in the regression procedure, we adjust the standard errors
using the procedure described by Newey and West (1994) to correct for
heteroscedasticity and serial correlation.
The results are graphically represented in Figure 2. The comovement
estimates using a 30-day estimation window are represented in panel A, whereas
those using the 90-day estimation window are reported in panel B.
Using the 30-day estimation window, we find that //optimistic decreases in the
StockDSKC rankings, whereas //pessimistic increases in the StockDSKC rankings,
as shown in panel A. To formally test our predictions, we calculate the difference
in ^pessimistic and ^optimistic across the highest and lowest StockDSKC quintile
groups. The difference is negative and statistically significant for //optimistic
(estimate = -0.071; t-value = -3.86) and is positive and statistically significant
for //pessimistic (estimate = 0.042; t-value = 2.08). The differences are slightly
larger for the //optimistic estimates. For the results from the 90-day estimation
window reported in panel B, //optimistic retains a similar pattern as before.
However, the difference is no longer statistically significant at the 10% level
(estimate = -0.020; t-value = -0.91). Likewise, for the //pessimistic estimates, the differences
remain positive but are also statistically insignificant (value =
0.014, f-value = 0.58).
In summary, the results provide evidence of return comovement attributable
to the investor mood proxy. They also suggest that the weather effect is not long
lived, given that weather patterns fluctuate more over longer time horizons.
Consequently, the mood effect related to weather is unlikely to persist for long
periods of time.
6. Conclusion
Several studies demonstrate that weather-based mood proxies explain variation
in trading volume and prices of broad-based stock indexes. They use cloud cover
in major stock market locations as a proxy for investor mood. In contrast to
these earlier studies, we use disaggregated data on the locations of institutional
investors to examine how weather-based measures of investor mood affect
perceptions and trading behavior, as well as individual stock returns. To our
knowledge, this study is the first to examine the direct impact of weather on
the trading behavior of institutional investors.31
We find that the weather-based mood measures affect institutional
investors' perceptions about market mispricing, thereby supporting mood-based explanations in
the existing literature. Specifically, using survey data, we
show that investor optimism, associated with lower values of deseasonalized
cloud cover, is negatively associated with perceived overpricing in both
individual stocks and the DJIA. Next, we use disaggregated trade data to
provide additional evidence that investor optimism increases their propensities
to buy. Finally, we also find that our stock-level investor mood measure
explains a significant amount of variation in daily returns of stocks associated
with higher arbitrage costs. In addition, we document return comovement
attributable to investor mood. Collectively, these findings complement existing
studies that document the effect of weather-induced mood on stock market
returns.
In future work, it would be interesting to examine whether other groups
of economic agents are also influenced by mood. For example, the decisions
of firm managers that have greater autonomy in decision-making, such as
those of small- and medium-sized enterprises, may be more susceptible to
mood-related biases, which, in turn, may significantly impact hiring and
investment decisions. The forecasts of sell-side equity analysts also may be
influenced by weather, which could affect how market participants incorporate
new information into prices. We leave these interesting questions for future research.